README.md

metadata

title: Programmable VLA
emoji: 🤖
colorFrom: purple
colorTo: blue
sdk: static
pinned: false
license: apache-2.0

Programmable VLA

RoboTwin 2.0 results: three standalone checkpoints across 50 tasks, a retained Kinex pilot, and the October 30 Kinex + VLAct Clean 50k run (1/3 native successes). Complete instructions, model inputs and outputs, sessions, usage, videos and artifacts are retained.

Standalone and agent results use different execution protocols and initializations. The previous standalone rosbridge pilot remains available in history.html.

Official videos show head, left wrist and right wrist at 4× the original 10 FPS recording. Wrist observations update at inference boundaries. This is action-index playback, not physical simulation time. Native 320×240 views are upscaled to the shared presentation layout.

The October 30 Kinex + VLA runs made 30 inferences, with 0 calls carrying additional structured conditions. The open interface supports caller-defined fields and formatting; observed use and effectiveness are separate questions. Previous valid episodes remain in Previous attempts.

VLAct All standalone evaluation

VLAct Qwen3-VL-OFT All (100K): 47/50 native successes, 50/50 valid completed episodes. The Base 40K standalone results and all Kinex episodes are retained.

Both standalone runs use the official RoboTwin evaluator, one admitted episode per task, RLE catalog instructions and native action limits. VLAct uses its released 32-action, image-bucket and wrap-joint contract. The current Kinex + VLA selection uses VLAct Clean 50k; its previous Base 40k episodes remain archived. Official expert admission selected different scene seeds in 7 of 50 tasks; the comparison is not fully paired. These results are not a reproduction of the upstream multi-trial paper score.

VLAct results · Protocol · Source identities

Evidence storage

This lightweight Space contains presentation code, logos, revision pointers and current instruction-audit snapshots. All evaluation result indexes, videos, images, trajectories, full sessions and historical evidence remain hosted in the paired Dataset, pinned by data-source.json. All existing pages and results are preserved.

VLAct Clean / Base standalone evaluation

StarVLA/VLAct_Qwen3OFT_Robotwin_Finetune, step 50,000: 37/50 native successes, 50/50 valid completed episodes. All 106 previously selected episodes are retained. The new result uses the same simulator, official admission rule, RLE instructions, native action limits and VLAct execution contract as the All run. Actual scene seeds agree with All in 45/50 tasks; training mixture and duration differ. This single-episode-per-task comparison is not the upstream paper protocol and does not isolate a causal mechanism.

Clean results · Protocol · Scene-seed comparison

RoboTwin 2.0 Hard · November 1, 2026

Open the Hard results: 50 tasks per condition, one valid episode each, native unseen instructions. Base 40k 2/50, VLAct Clean 16/50, VLAct All 46/50, pure Kinex Astra/high 33/50. The existing Easy and historical panels are retained. Execution interfaces, seeds and physical budgets differ; the panel documents those limits.

RoboTwin 2.0-Plus · November 2, 2026

Open the Plus results: 1,200 valid episodes, three checkpoints × 50 tasks × eight configurations × one admitted episode, with native instructions. Base 40k 94/400 (23.5%), VLAct Clean 50k 222/400 (55.5%), VLAct All 100k 294/400 (73.5%). All episodes have native videos and full evidence ZIPs. Existing Easy, Hard and historical pages are retained.

Single-episode coverage is not the upstream multi-trial protocol; vision-noise index 0 covers motion noise only. The page and report document expert admission, native language pools, checkpoint identities and the All recovery. No valid outcomes were retried. Dataset revision: 9f8a68a40148e1ad013b35a2d8774ff113cb0c5e.

Unified Plus presentation · November 2, 2026

The Plus panel now uses the Easy/Hard task browser, comparison table, episode tabs, video player, evidence panels and VLA input/output viewer. All 1,200 videos use the same head + left wrist + right wrist layout, 2880 × 720 output, H.264 CRF 20, 40 FPS / 4× action-index playback, posters and last-frame controls. Native 320 × 240 camera views are upscaled; wrist observations update at inference boundaries. No new observations or benchmark outcomes were generated.

All original scores (94/400, 222/400, 294/400), native videos, evidence ZIPs, seeds and instructions are unchanged. VLA I/O includes every recorded call and byte-identical input/source PNGs. Presentation protocol. Dataset revision: b2d67a67dd2f9d8e12c1f8db13ba8de309739883.

Non-VLAct OFT All 140k

The Easy, Hard and Plus panels include StarVLA/Qwen3-VL-OFT-RoboTwin2-All (140k; 27,500 clean/randomized training demonstrations; 50-action chunks). Each panel preserves its established instruction and native evaluation protocol. Completion counts are reported separately from successes. All prior episodes remain unchanged.

Campaign evidence.

Current instruction audits · November 3, 2026

Easy · Hard · Plus

The audit compares the current RLE-Bench-inhouse instruction templates with pinned upstream full_description text, using the reference benchmark's side-by-side word diff. Search, configuration, changed/unchanged and scene-bound filters cover all 50 Easy tasks, 50 Hard tasks and 400 Plus task/configuration pairs. Exact-text differences affect 45/50, 49/50 and 400/400 entries respectively. These counts use full_description as the baseline; they are not counts of changed evaluation outcomes.

Expandable panels show the corresponding native runtime language pools, including Plus R1/R2/R3, and the instructions recorded for published episodes. RLE scene bindings, pinned native sources and verifier links are included. The RLE language-plus preset uses the explicit task goal and removes the original language perturbation; it is a separate instruction protocol.

The current catalog snapshots include local, uncommitted RLE instruction revisions; source hashes, YAML snapshots and effective audit JSON make this snapshot reproducible. This is a dated snapshot, not a live mirror of the source checkout. The Hard and Plus published results retain their native-language protocol. The audit does not relabel any historical episode or change Dataset revisions, scores, videos, costs or evidence.

Kinex + VLAct Qwen3-VL-OFT · Clean 50k

Complete: 100/100 audited episodes published. Updated: 2026-11-06T21:02:24.441645+00:00.

• Agent: Kinex, GPT-6 Astra / high.
• VLA checkpoint: StarVLA/VLAct_Qwen3OFT_Robotwin_Finetune, step 50,000, RoboTwin clean training distribution.
• Frozen checkpoint revision: dcf0f1d7239b28ae81a039c8d05bdfcd1d4792a5.
• Weight file: checkpoints/steps_50000_pytorch_model.pt.
• Weight SHA-256: e3823ca683057371ab94d4d66434538e2e2a53b861a87ee77a1981f941209873.
• Benchmark: RoboTwin 2.0 Hard, 50 tasks per protocol, one authored seed per task.

Distinct result entry	Input protocol
Kinex + VLAct Qwen3-VL-OFT · Clean 50k (autonomous)	verbatim-text-v1: agent-written free text.
Kinex + VLAct Qwen3-VL-OFT · Clean 50k (input-required · numeric targets)	numeric-target-text-v1: instruction plus one or two numerical TCP/joint targets; optional feedback. Null or qualitative targets are rejected.

Both protocols let Kinex choose whether to call VLA and how to execute its output. Zero VLA calls are valid. Numerical targets enter the frozen VLA as language text; validation checks their structure, not grounding, reachability or learned compliance.

This campaign is separate from Kinex + Qwen3-VL-OFT · All 140k (input-required · text targets), whose historical protocol requires five JSON fields and allows qualitative targets or null; numerical targets are optional. It is also separate from standalone VLAct Clean 50k, which uses a different agent/execution interface. Existing results and evidence remain intact.

Protocol	Published / planned	Successes / evaluated	Episodes using VLA
autonomous	50 / 50	31 / 50	47
input-required	50 / 50	43 / 50	4

Pending tasks are not failures and are excluded from success-rate denominators. Partial results are not a full-suite score. All completed valid outcomes are retained; no success-based selection or valid-outcome retries.

VLA input: three RGB images and caller text. Output: 32 absolute-joint samples at 60 ms, resampled to 40 ms controls; execution prefixes are chosen by Kinex. No fine-tuning is performed in this evaluation.

Kinex + VLAct Qwen3-VL-OFT · Clean 50k (input-required · text targets)

Complete: 50/50 audited Hard episodes published. Updated: 2026-11-06T22:17:27.675386+00:00.

• Agent: Kinex, GPT-6 Astra / high.
• Checkpoint: StarVLA/VLAct_Qwen3OFT_Robotwin_Finetune, step 50,000, RoboTwin clean training distribution.
• Frozen weight SHA-256: e3823ca683057371ab94d4d66434538e2e2a53b861a87ee77a1981f941209873.
• Input protocol: input-required · text targets, contract text-targets-v1.
• Benchmark: RoboTwin 2.0 Hard, all 50 authored task/seed/budget combinations; one fresh episode each.

Every VLA request serializes a JSON object with goal, context, feedback, constraints and target. All five fields are required; qualitative targets, grounded numerical values, null and additional fields are allowed. Missing fields, duplicate keys, nonfinite numbers and non-object JSON are rejected. The full caller-written text enters the frozen VLA unchanged. Neither numerical encoding nor target enforcement is added.

Kinex retains control of observations, whether to invoke VLA, direct control, execution prefixes and feedback. Zero VLA calls are valid. This uses the historical five-field text-target schema within the current shared operation description and model/execution pipeline. The present numeric-target-text-v1 results require numerical TCP/joint destinations; the two input-required variants remain separate. Earlier All 140k text-target results retain their own checkpoint and frozen runtime/prompt versions.

Native successes: 29/50 evaluated. Episodes using VLA: 43; zero-call episodes: 7. VLA calls: 232; rejected requests: 0; accepted calls with null target: 0.

Pending tasks are excluded from score denominators. Partial results are not full-suite scores. All valid successes and failures are retained; no valid-outcome retries, seed substitution, fine-tuning or expert-success filtering. One authored seed per task does not establish multi-seed reliability.

The released VLAct model uses three RGB images, aspect-ratio buckets 320 x 180 / 280 x 210, 32 absolute-joint samples at 60 ms and wrap decoding. Kinex selects prefixes resampled to 40 ms controls. Existing authored instructions, initialization seeds, model seed 42 and native step limits match the preceding Clean 50k campaign. Local GPU sharing can vary, so latency is not a controlled comparison.

API-equivalent costs cover reported-response usage. Failed requests without terminal usage remain unknown. Account billing and local GPU costs are not measured.